NVIDIA Adds a $4,999 DGX Spark With 64GB of Memory for Local AI

NVIDIA is giving developers another starting point for running AI without a cloud instance: a 64GB DGX Spark system starting at $4,999. Announced October 2, the configuration is scheduled to reach six hardware partners on October 23, 2026, with a path to connect a second machine as workloads grow.

Acer, ASUS, Dell, Gigabyte, HP and MSI will sell the new configuration, which is available exclusively through manufacturer partners. It retains the GB10 Grace Blackwell Superchip, DGX OS and full NVIDIA AI software stack used by the 128GB model. The change is a new memory configuration, not a different computing platform.

One machine for models and everyday development

The new configuration is designed to run models and agents entirely on the device, without a cloud dependency. NVIDIA says a single 64GB system supports models with up to 100 billion parameters. That is a company-stated capacity ceiling, rather than a measured speed result for every model within that size.

DGX Spark combines its processor and unified memory with ConnectX-7 networking and NVIDIA’s CUDA-accelerated software. NVIDIA positions the system for inference—running a model—as well as fine-tuning, data science and edge development. Its supported software includes NVIDIA Agent Toolkit, CUDA-X AI libraries, Nemotron models, Ollama, vLLM and PyTorch with CUDA.

The intended workflows include keeping a coding or research agent running around the clock, reviewing code, analyzing documents and handling multistep tasks. Another option is to run a language or image model on Spark while using an application on a laptop or desktop. NVIDIA says that arrangement leaves the everyday PC free for other work.

When the workload needs a second Spark

The next step is a two-machine cluster: two systems connected to work together. Each Spark has a built-in ConnectX-7 network adapter. A direct QSFP cable connection lets two 64GB units pool their memory to 128GB over a 200-gigabit Ethernet connection. NVIDIA says the pair can support models up to 200 billion parameters.

Sync Cluster Assistant handles the connection’s software setup. It detects the attached units, checks their configuration and configures the ConnectX-7 network. Because both machines run the same software stack, NVIDIA says developers do not need to reconfigure their software environment when moving from one unit to two.

The performance evidence is narrower than the capacity claim. In NVIDIA’s Qwen 3.8 27B test, two clustered 64GB systems delivered up to 1.7 times the performance of one. The result is tied to that company-run test; it is not a promise that adding a second machine will double speed across workloads.

Manual setup first, a launcher later in October

For developers getting started, NVIDIA outlines a download-and-connect workflow rather than requiring the forthcoming launcher:

1. Download a supported model-running framework, such as llama.cpp, Ollama, vLLM or LM Studio.
2. Download the recommended local model for the intended workflow.
3. For two systems, connect their ConnectX-7 ports and launch Sync Cluster Assistant to configure the network and route workloads.

NVIDIA Sync Model Launcher is planned for the end of October. It will download and launch Qwen3.8 27B on one Spark or a cluster, configure the model across connected devices and make it accessible from a laptop. NVIDIA also says the launcher will configure OpenCode so developers can begin coding in their browser.